Lichang Chen*, Shiyang Li*, Jun Yan, Hai Wang, Kalpa Gunaratna, Vikas Yadav, Zheng Tang, Vijay Srinivasan, Tianyi Zhou, Heng Huang, Hongxia Jin
Our Model "AlpaGasus"is pronounced as "/ˈælpəˈɡeɪsəs/", or "/ˈælpəˈɡəsəs/". The logo is generated by Midjourney
- [2023.7] We release our paper. If you have any questions about our project, please send email to bobchen@umd.edu
- [2023.9] Thanks to @GPT4animal for reimplementing the results in our paper. Please check this fantastic repo: https://github.com/gpt4life/alpagasus.
- [2023.9] Thanks to @gauss5930 and @YooYunS who implemented the QLoRA version of Alpagasus-7B and 13B, which could be run on the customer-level GPUs. please refer to their repo: Alpagasus2-QLoRA They also show that tuning LLaMA-2 could achieve better performance.
If you find our paper useful, please consider citing:
@inproceedings{
chen2024alpagasus,
title={AlpaGasus: Training a Better Alpaca with Fewer Data},
author={Lichang Chen and Shiyang Li and Jun Yan and Hai Wang and Kalpa Gunaratna and Vikas Yadav and Zheng Tang and Vijay Srinivasan and Tianyi Zhou and Heng Huang and Hongxia Jin},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=FdVXgSJhvz}
}